AI Models in Auto Learning Agents

The platform's central design idea is that models are replaceable workers inside a stable structure. Instructions live in editable files, context comes from the conversation system, memory persists in the memory bank, and abilities come from the tool layer. Because all of that lives outside the model, you can change which model an agent thinks with as easily as changing a form field, and the agent carries on with the same knowledge, history, and behavior.

The Model Keys

One set of keys is used everywhere: the chat model picker, each agent's configuration, and each pipeline's settings.

Claude. claude-sonnet is the fast, economical default for conversational and routine work. claude-opus-high, claude-opus-xhigh, and claude-opus-max run Claude Opus at increasing effort levels, the dial for how much thinking a task deserves. The pipelines that plan, build, and review code benefit most from the higher levels.

OpenAI Codex. codex-mini runs GPT-5.4 Mini for quick, inexpensive work. codex-5.5, codex-5.5-high, and codex-5.5-xhigh run GPT-5.5 at increasing effort levels, mirroring the Claude ladder so you can compare like for like.

Ollama. ollama-<modelname> addresses any model you serve through a local Ollama instance, ollama-llama3, ollama-qwen3, whatever you have pulled. Local models turn high-frequency background agents free to run, and they keep every token on your own hardware.

The models_installed list in settings.txt declares which families this install uses and filters what every model picker offers, so a Claude-only install never shows options it cannot run. Mixed installs are fully supported and common.

How Model Calls Run

Model access is built on the providers' own agent tooling, which is what makes subscription accounts usable alongside API keys. Claude calls run headless through the claude CLI with a system-minted session id, the acceptEdits permission mode, and an allowlist that covers the platform's tool directory and web tools, so agents can act without a human at the keyboard while staying inside defined boundaries. Codex calls run in a workspace-write sandbox with network access enabled, the equivalent arrangement on the OpenAI side.

Multi-step work gets a powerful guarantee: one continuous session per task. When the coding pipeline or the research pipeline runs a task through its passes, the first step creates the session and every later step resumes it, so the model that reviews the code is the same session that planned and wrote it, with the full history in context. The one-shot coding and research tools work the same way. This is why multi-pass output reads like one engineer's work rather than a relay race.

Connecting Your Accounts

Each family connects in the way its provider works best. Claude uses the claude CLI login, run claude login once (inside the container on Docker installs) and the credential persists across restarts and upgrades, which lets a Claude subscription power the whole platform. Codex connects through its own CLI setup on the same machine. Ollama needs no credentials at all, point it at your local instance and pull the models you want. Whichever mix you connect, declare it in models_installed and every picker in the UI stays in sync.

Choosing Models per Agent

The practical strategy is to match the model to the job and let the structure do the rest. Conversational work and routine processing run beautifully on claude-sonnet or codex-mini. Agents that compose customer-facing text, like the marketing and support presets, deserve a mid-tier model. The coding pipeline earns the top of the ladder, the difference shows directly in code quality, and a high effort level costs less than a human review cycle.

Cost has two dials per agent: the model key and the tick rate. A background agent on a cheap model with a relaxed tick costs close to nothing; the same agent pointed at Opus on a one-minute tick is a very different bill. Start cheap, watch the results in the activity feed, and promote only the agents whose work improves with a stronger model. Every agent can also be paused individually, which makes experimentation reversible. The creating agents guide covers where these fields live.

Using Several Models Together

Two tools make multi-model setups more than the sum of their parts. The mixture tool sends the same question to several models at once and synthesizes the answers into one response, useful for decisions where a second opinion matters. The delegation tool spawns a focused child agent with its own isolated context and its own model choice, so an expensive model can hand routine subtasks to a cheap one. Both are ordinary tools, available to any agent, and the tools guide covers them alongside the rest of the layer.

Training Your Own

Installs that run local models can go one step further: with trainOllama enabled in settings, the schedule node runs a weekly fine-tuning pass that trains a local model on your own memory data, the accumulated knowledge, procedures, and outcomes your system has gathered. The result is a local model that speaks your install's language. The machine learning guide covers the full learning layer this sits inside.